GRIP: Grounded Reasoning via Information-Restricted Premises
2026-08-17 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors identify a problem called "query dominance" where the query itself overshadows the useful information from retrieved evidence during reasoning tasks. To fix this, they created GRIP, a system that limits how much of the retrieved evidence the model can directly use, forcing it to focus on new information not already in the query. Testing on five reasoning tasks, their method worked better than other strong models, showed much less overlap between the query and evidence, and reduced mistakes in generating information. They also found that the evidence representations in GRIP are more distinct from the query than in baseline models.
retrieval-augmented generationquery dominancelatent statestochastic bottleneckmutual informationhallucination reductionresidual-alignmentreasoning benchmarksdecoderencoder
Authors
Lirui Teng
Abstract
High-capacity encoders in retrieval-augmented generation (RAG) can let the query dominate the latent state, leaving retrieved evidence functionally irrelevant. We call this failure mode query dominance. To address it, we introduce \textbf{GRIP} (Grounded Reasoning via Information-Restricted Premises), which imposes capacity asymmetry: the decoder keeps full-dimensional access to the query, while retrieved evidence passes through a severe stochastic bottleneck. This forces the evidence channel to encode only the residual information unavailable from the query. Across five reasoning benchmarks, GRIP outperforms strong iterative baselines, cuts a query--latent mutual-information diagnostic by roughly 30$\times$ (14.8 $\to$ 0.47 bits), and reduces hallucination by 73\%. Residual-alignment analysis further shows that the bottleneck output occupies subspaces less aligned with the query than baseline representations.